Self-Contrastive Graph Diffusion Clustering for Hyperspectral Images

Hyperspectral image (HSI) clustering is widely utilized in remote sensing applications, yet it remains a challenging unsupervised learning task. In current deep HSI clustering frameworks, the concept of contrastive learning is widely adopted. However, there are two main issues: on the one hand, researchers need to carefully design data augmentation techniques to reduce noise introduction; on the other hand, traditional contrastive learning methods are limited to selecting negative samples under random strategies, which may mistakenly identify semantically consistent samples as negative pairs. Both issues affect the clustering performance. To address these challenges, we introduce a contrastive learning-based clustering framework, termed self-contrastive graph diffusion clustering, which effectively integrates pretraining techniques with graph diffusion networks for enhanced clustering performance. Specifically, we first conduct pretraining at the pixel level for HSI to extract its multilevel spatial and spectral features. Subsequently, we conduct entropy rate superpixel segmentation, build a superpixel graph, and shift the clustering process from pixel-level modeling to superpixel-level graph learning. Then, we adopt the attention module to aggregate the structural and feature information, thereby obtaining high-quality embeddings. Next, we use the diffusion module to balance node states in the graph. Regarding the sampling strategy, we comprehensively consider the structural and feature information to obtain high-quality positive and negative sample pairs. We ensure that two nodes are regarded as a negative sample pair only when they are not spatially connected and are also unrelated in the k-nearest neighbors graph. Experiments show that the proposed model achieves higher clustering metrics than existing methods on all three datasets, namely Salinas, Pavia University, and Trento.